Both qemu and qemu-img use writeback cache mode by default, which is
already documented in qemu(1). qemu-nbd uses writethrough cache mode by
default, and the default cache mode is not documented.
According to the qemu-nbd(8):
--cache=CACHE
The cache mode to be used with the file. See the
documentation of the emulator's -drive cache=... option for
allowed values.
qemu(1) says:
The default mode is cache=writeback.
So users have no reason to assume that qemu-nbd is using writethough
cache mode. The only hint is the painfully slow writing when using the
defaults.
Looking in git history, it seems that qemu used writethrough in the past
to support broken guests that did not flush data properly, or could not
flush due to limitations in qemu. But qemu-nbd clients can use
NBD_CMD_FLUSH to flush data, so using writethrough does not help anyone.
Change the default cache mode to writback, and document the default and
available values properly in the online help and manual.
With this change converting image via qemu-nbd is 3.5 times faster.
$ qemu-img create dst.img 50g
$ qemu-nbd -t -f raw -k /tmp/nbd.sock dst.img
Before this change:
$ hyperfine -r3 "./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock"
Benchmark #1: ./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock
Time (mean ± σ): 83.639 s ± 5.970 s [User: 2.733 s, System: 6.112 s]
Range (min … max): 76.749 s … 87.245 s 3 runs
After this change:
$ hyperfine -r3 "./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock"
Benchmark #1: ./qemu-img convert -p -f raw -O raw -T none -W fedora34.img nbd+unix:///?socket=/tmp/nbd.sock
Time (mean ± σ): 23.522 s ± 0.433 s [User: 2.083 s, System: 5.475 s]
Range (min … max): 23.234 s … 24.019 s 3 runs
Users can avoid the issue by using --cache=writeback[1] but the defaults
should give good performance for the common use case.
[1] https://bugzilla.redhat.com/1990656
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Message-Id: <20210813205519.50518-1-nsoffer@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
CC: qemu-stable@nongnu.org
Signed-off-by: Eric Blake <eblake@redhat.com>
When querying image extents for raw image, qemu-nbd reports holes as
zero:
$ qemu-nbd -t -r -f raw empty-6g.raw
$ qemu-img map --output json nbd://localhost
[{ "start": 0, "length": 6442450944, "depth": 0, "zero": true, "data": true, "offset": 0}]
$ qemu-img map --output json empty-6g.raw
[{ "start": 0, "length": 6442450944, "depth": 0, "zero": true, "data": false, "offset": 0}]
Turns out that qemu-img map reports a hole based on BDRV_BLOCK_DATA, but
nbd server reports a hole based on BDRV_BLOCK_ALLOCATED.
The NBD protocol says:
NBD_STATE_HOLE (bit 0): if set, the block represents a hole (and
future writes to that area may cause fragmentation or encounter an
NBD_ENOSPC error); if clear, the block is allocated or the server
could not otherwise determine its status.
qemu-img manual says:
whether the sectors contain actual data or not (boolean field data;
if false, the sectors are either unallocated or stored as
optimized all-zero clusters);
To me, data=false looks compatible with NBD_STATE_HOLE. From user point
of view, getting same results from qemu-nbd and qemu-img is more
important than being more correct about allocation status.
Changing nbd server to report holes using BDRV_BLOCK_DATA makes qemu-nbd
results compatible with qemu-img map:
$ qemu-img map --output json nbd://localhost
[{ "start": 0, "length": 6442450944, "depth": 0, "zero": true, "data": false, "offset": 0}]
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Message-Id: <20210219160752.1826830-1-nsoffer@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
FilePath creates now one temporary file:
with FilePath("a") as a:
Or more:
with FilePath("a", "b", "c") as (a, b, c):
This is also the behavior of the file_path() helper, used by some of the
tests. Now we have only 2 helpers for creating temporary files instead
of 3.
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-Id: <20200828232152.205833-5-nsoffer@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Accept variable number of names instead of a sequence:
with FilePaths("a", "b", "c") as (a, b, c):
The disadvantage is that base_dir must be used as kwarg:
with FilePaths("a", "b", base_dir=soc_dir) as (sock1, sock2):
But this is more clear and calling optional argument as positional
arguments is bad idea anyway.
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-Id: <20200828232152.205833-4-nsoffer@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
When this class was extracted from FilePath, the docstring was not
updated for generating multiple files, and the example usage was
referencing unrelated file.
While fixing the docstring, add example for creating sockets, which
should use iotests.sock_dir instead of the default base_dir.
Fixes: de263986b5
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Message-Id: <20200828232152.205833-3-nsoffer@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Add 2 helpers for measuring and checking images:
- qemu_img_measure()
- qemu_img_check()
Both use --output-json and parse the returned json to make easy to use
in other tests. I'm going to use them in a new test, and I hope they
will be useful in may other tests.
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Message-Id: <20200727215846.395443-4-nsoffer@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
Instead of duplicating the code to wait until the server is ready and
remember to terminate the server and wait for it, make it possible to
use like this:
with qemu_nbd_popen('-k', sock, image):
# Access image via qemu-nbd socket...
Only test 264 used this helper, but I had to modify the output since it
did not consistently when starting and stopping qemu-nbd.
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Message-Id: <20200727215846.395443-3-nsoffer@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Signed-off-by: Eric Blake <eblake@redhat.com>
When converting to qcow2 compressed format, the last step is a special
zero length compressed write, ending in a call to bdrv_co_truncate(). This
call always fails for the nbd driver since it does not implement
bdrv_co_truncate().
For block devices, which have the same limits, the call succeeds since
the file driver implements bdrv_co_truncate(). If the caller asked to
truncate to the same or smaller size with exact=false, the truncate
succeeds. Implement the same logic for nbd.
Example failing without this change:
In one shell start qemu-nbd:
$ truncate -s 1g test.tar
$ qemu-nbd --socket=/tmp/nbd.sock --persistent --format=raw --offset 1536 test.tar
In another shell convert an image to qcow2 compressed via NBD:
$ echo "disk data" > disk.raw
$ truncate -s 1g disk.raw
$ qemu-img convert -f raw -O qcow2 -c disk1.raw nbd+unix:///?socket=/tmp/nbd.sock; echo $?
1
qemu-img failed, but the conversion was successful:
$ qemu-img info nbd+unix:///?socket=/tmp/nbd.sock
image: nbd+unix://?socket=/tmp/nbd.sock
file format: qcow2
virtual size: 1 GiB (1073741824 bytes)
...
$ qemu-img check nbd+unix:///?socket=/tmp/nbd.sock
No errors were found on the image.
1/16384 = 0.01% allocated, 100.00% fragmented, 100.00% compressed clusters
Image end offset: 393216
$ qemu-img compare disk.raw nbd+unix:///?socket=/tmp/nbd.sock
Images are identical.
Fixes: https://bugzilla.redhat.com/1860627
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Message-Id: <20200727215846.395443-2-nsoffer@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
[eblake: typo fixes]
Signed-off-by: Eric Blake <eblake@redhat.com>
Replace instances of:
(n & (BDRV_SECTOR_SIZE - 1)) == 0
And:
(n & ~BDRV_SECTOR_MASK) == 0
With:
QEMU_IS_ALIGNED(n, BDRV_SECTOR_SIZE)
Which reveals the intent of the code better, and makes it easier to
locate the code checking alignment.
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Message-id: 20190827185913.27427-2-nsoffer@redhat.com
Reviewed-by: John Snow <jsnow@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Using block_resize we can test allocate_first_block() with file
descriptor opened with O_DIRECT, ensuring that it works for any size
larger than 4096 bytes.
Testing smaller sizes is tricky as the result depends on the filesystem
used for testing. For example on NFS any size will work since O_DIRECT
does not require any alignment.
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-id: 20190827010528.8818-3-nsoffer@redhat.com
Signed-off-by: Max Reitz <mreitz@redhat.com>
When creating an image with preallocation "off" or "falloc", the first
block of the image is typically not allocated. When using Gluster
storage backed by XFS filesystem, reading this block using direct I/O
succeeds regardless of request length, fooling alignment detection.
In this case we fallback to a safe value (4096) instead of the optimal
value (512), which may lead to unneeded data copying when aligning
requests. Allocating the first block avoids the fallback.
Since we allocate the first block even with preallocation=off, we no
longer create images with zero disk size:
$ ./qemu-img create -f raw test.raw 1g
Formatting 'test.raw', fmt=raw size=1073741824
$ ls -lhs test.raw
4.0K -rw-r--r--. 1 nsoffer nsoffer 1.0G Aug 16 23:48 test.raw
And converting the image requires additional cluster:
$ ./qemu-img measure -f raw -O qcow2 test.raw
required size: 458752
fully allocated size: 1074135040
When using format like vmdk with multiple files per image, we allocate
one block per file:
$ ./qemu-img create -f vmdk -o subformat=twoGbMaxExtentFlat test.vmdk 4g
Formatting 'test.vmdk', fmt=vmdk size=4294967296 compat6=off hwversion=undefined subformat=twoGbMaxExtentFlat
$ ls -lhs test*.vmdk
4.0K -rw-r--r--. 1 nsoffer nsoffer 2.0G Aug 27 03:23 test-f001.vmdk
4.0K -rw-r--r--. 1 nsoffer nsoffer 2.0G Aug 27 03:23 test-f002.vmdk
4.0K -rw-r--r--. 1 nsoffer nsoffer 353 Aug 27 03:23 test.vmdk
I did quick performance test for copying disks with qemu-img convert to
new raw target image to Gluster storage with sector size of 512 bytes:
for i in $(seq 10); do
rm -f dst.raw
sleep 10
time ./qemu-img convert -f raw -O raw -t none -T none src.raw dst.raw
done
Here is a table comparing the total time spent:
Type Before(s) After(s) Diff(%)
---------------------------------------
real 530.028 469.123 -11.4
user 17.204 10.768 -37.4
sys 17.881 7.011 -60.7
We can see very clear improvement in CPU usage.
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Message-id: 20190827010528.8818-2-nsoffer@redhat.com
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
In some cases buf_align or request_alignment cannot be detected:
1. With Gluster, buf_align cannot be detected since the actual I/O is
done on Gluster server, and qemu buffer alignment does not matter.
Since we don't have alignment requirement, buf_align=1 is the best
value.
2. With local XFS filesystem, buf_align cannot be detected if reading
from unallocated area. In this we must align the buffer, but we don't
know what is the correct size. Using the wrong alignment results in
I/O error.
3. With Gluster backed by XFS, request_alignment cannot be detected if
reading from unallocated area. In this case we need to use the
correct alignment, and failing to do so results in I/O errors.
4. With NFS, the server does not use direct I/O, so both buf_align cannot
be detected. In this case we don't need any alignment so we can use
buf_align=1 and request_alignment=1.
These cases seems to work when storage sector size is 512 bytes, because
the current code starts checking align=512. If the check succeeds
because alignment cannot be detected we use 512. But this does not work
for storage with 4k sector size.
To determine if we can detect the alignment, we probe first with
align=1. If probing succeeds, maybe there are no alignment requirement
(cases 1, 4) or we are probing unallocated area (cases 2, 3). Since we
don't have any way to tell, we treat this as undetectable alignment. If
probing with align=1 fails with EINVAL, but probing with one of the
expected alignments succeeds, we know that we found a working alignment.
Practically the alignment requirements are the same for buffer
alignment, buffer length, and offset in file. So in case we cannot
detect buf_align, we can use request alignment. If we cannot detect
request alignment, we can fallback to a safe value. To use this logic,
we probe first request alignment instead of buf_align.
Here is a table showing the behaviour with current code (the value in
parenthesis is the optimal value).
Case Sector buf_align (opt) request_alignment (opt) result
======================================================================
1 512 512 (1) 512 (512) OK
1 4096 512 (1) 4096 (4096) FAIL
----------------------------------------------------------------------
2 512 512 (512) 512 (512) OK
2 4096 512 (4096) 4096 (4096) FAIL
----------------------------------------------------------------------
3 512 512 (1) 512 (512) OK
3 4096 512 (1) 512 (4096) FAIL
----------------------------------------------------------------------
4 512 512 (1) 512 (1) OK
4 4096 512 (1) 512 (1) OK
Same cases with this change:
Case Sector buf_align (opt) request_alignment (opt) result
======================================================================
1 512 512 (1) 512 (512) OK
1 4096 4096 (1) 4096 (4096) OK
----------------------------------------------------------------------
2 512 512 (512) 512 (512) OK
2 4096 4096 (4096) 4096 (4096) OK
----------------------------------------------------------------------
3 512 4096 (1) 4096 (512) OK
3 4096 4096 (1) 4096 (4096) OK
----------------------------------------------------------------------
4 512 4096 (1) 4096 (1) OK
4 4096 4096 (1) 4096 (1) OK
I tested that provisioning VMs and copying disks on local XFS and
Gluster with 4k bytes sector size work now, resolving bugs [1],[2].
I tested also on XFS, NFS, Gluster with 512 bytes sector size.
[1] https://bugzilla.redhat.com/1737256
[2] https://bugzilla.redhat.com/1738657
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
With Kevin's "block: Fix slow pre-zeroing in qemu-img convert"[1]
(commit c9fdcf202f, 'qemu-img: Use BDRV_REQ_NO_FALLBACK for
pre-zeroing') we skip the pre zero step called like this:
blk_make_zero(s->target, BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK)
And we write zeroes later using:
blk_co_pwrite_zeroes(s->target,
sector_num << BDRV_SECTOR_BITS,
n << BDRV_SECTOR_BITS, 0);
Since we use flags=0, this is translated to NBD_CMD_WRITE_ZEROES with
NBD_CMD_FLAG_NO_HOLE flag, which cause the NBD server to allocated space
instead of punching a hole.
Here is an example failure:
$ dd if=/dev/urandom of=src.img bs=1M count=5
$ truncate -s 50m src.img
$ truncate -s 50m dst.img
$ nbdkit -f -v -e '' -U nbd.sock file file=dst.img
$ ./qemu-img convert -n src.img nbd:unix:nbd.sock
We can see in nbdkit log that it received the NBD_CMD_FLAG_NO_HOLE
(may_trim=0):
nbdkit: file[1]: debug: newstyle negotiation: flags: export 0x4d
nbdkit: file[1]: debug: pwrite count=2097152 offset=0
nbdkit: file[1]: debug: pwrite count=2097152 offset=2097152
nbdkit: file[1]: debug: pwrite count=1048576 offset=4194304
nbdkit: file[1]: debug: zero count=33554432 offset=5242880 may_trim=0
nbdkit: file[1]: debug: zero count=13631488 offset=38797312 may_trim=0
nbdkit: file[1]: debug: flush
And the image became fully allocated:
$ qemu-img info dst.img
virtual size: 50M (52428800 bytes)
disk size: 50M
With this change we see that nbdkit did not receive the
NBD_CMD_FLAG_NO_HOLE (may_trim=1):
nbdkit: file[1]: debug: newstyle negotiation: flags: export 0x4d
nbdkit: file[1]: debug: pwrite count=2097152 offset=0
nbdkit: file[1]: debug: pwrite count=2097152 offset=2097152
nbdkit: file[1]: debug: pwrite count=1048576 offset=4194304
nbdkit: file[1]: debug: zero count=33554432 offset=5242880 may_trim=1
nbdkit: file[1]: debug: zero count=13631488 offset=38797312 may_trim=1
nbdkit: file[1]: debug: flush
And the file is sparse as expected:
$ qemu-img info dst.img
virtual size: 50M (52428800 bytes)
disk size: 5.0M
[1] http://lists.nongnu.org/archive/html/qemu-block/2019-03/msg00761.html
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Now that we are truncating the file in both PREALLOC_MODE_FULL and
PREALLOC_MODE_OFF, not truncating in PREALLOC_MODE_FALLOC looks odd.
Add a comment explaining why we do not truncate in this case.
Signed-off-by: Nir Soffer <nirsof@gmail.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
In a previous commit (qemu-img: Do not truncate before preallocation) we
moved truncate to the PREALLOC_MODE_OFF branch to avoid slowdown in
posix_fallocate().
However this change is not optimal when using PREALLOC_MODE_FULL, since
knowing the final size from the beginning could allow the file system
driver to do less allocations and possibly avoid fragmentation of the
file.
Now we truncate also before doing full preallocation.
Signed-off-by: Nir Soffer <nirsof@gmail.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Add tests for creating raw image with and without the preallocation
option.
Signed-off-by: Nir Soffer <nirsof@gmail.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
When using file system that does not support fallocate() (e.g. NFS <
4.2), truncating the file only when preallocation=OFF speeds up creating
raw file.
Here is example run, tested on Fedora 24 machine, creating raw file on
NFS version 3 server.
$ time ./qemu-img-master create -f raw -o preallocation=falloc mnt/test 1g
Formatting 'mnt/test', fmt=raw size=1073741824 preallocation=falloc
real 0m21.185s
user 0m0.022s
sys 0m0.574s
$ time ./qemu-img-fix create -f raw -o preallocation=falloc mnt/test 1g
Formatting 'mnt/test', fmt=raw size=1073741824 preallocation=falloc
real 0m11.601s
user 0m0.016s
sys 0m0.525s
$ time dd if=/dev/zero of=mnt/test bs=1M count=1024 oflag=direct
1024+0 records in
1024+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 15.6627 s, 68.6 MB/s
real 0m16.104s
user 0m0.009s
sys 0m0.220s
Running with strace we can see that without this change we do one
pread() and one pwrite() for each block. With this change, we do only
one pwrite() per block.
$ strace ./qemu-img-master create -f raw -o preallocation=falloc mnt/test 8192
...
pread64(9, "\0", 1, 4095) = 1
pwrite64(9, "\0", 1, 4095) = 1
pread64(9, "\0", 1, 8191) = 1
pwrite64(9, "\0", 1, 8191) = 1
$ strace ./qemu-img-fix create -f raw -o preallocation=falloc mnt/test 8192
...
pwrite64(9, "\0", 1, 4095) = 1
pwrite64(9, "\0", 1, 8191) = 1
This happens because posix_fallocate is checking if each block is
allocated before writing a byte to the block, and when truncating the
file before preallocation, all blocks are unallocated.
Signed-off-by: Nir Soffer <nirsof@gmail.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
The result of openfile was not checked, leading to failure deep in the
actual command with confusing error message, and exiting with exit code 0.
Here is a simple example - trying to read with the wrong format:
$ touch file
$ qemu-io -f qcow2 -c 'read -P 1 0 1024' file; echo $?
can't open device file: Image is not in qcow2 format
no file open, try 'help open'
0
With this patch, we fail earlier with exit code 1:
$ ./qemu-io -f qcow2 -c 'read -P 1 0 1024' file; echo $?
can't open device file: Image is not in qcow2 format
1
Failing earlier, we don't log this error now:
no file open, try 'help open'
But some tests expected it; the line was removed from the test output.
Signed-off-by: Nir Soffer <nirsof@gmail.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Message-id: 20170201003120.23378-2-nirsof@gmail.com
Reviewed-by: Max Reitz <mreitz@redhat.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>
Nir Soffer